When making videos, the most difficult images to generate are often not bright, clean product shots, but dark scenes such as night scenes, disasters, science fiction, and war.
These kinds of shots may look “dark,” but they contain no less information at all.
There need to be brightness variations at different depths in the clouds, smoke cannot blur into a single mass, distant buildings need to maintain spatial layers, flames and energy light sources need to illuminate the surrounding environment, and rain, metal, and standing water also need to form reasonable reflections. When characters wear dark clothing, their outlines must neither disappear into the background nor float in the frame like stickers.
Images with insufficient quality may feel very atmospheric when reduced, but once they enter the video production workflow, problems will quickly be exposed. Color banding appears in dark areas, building details blend into one another, smoke loses its layers, aircraft structures become blurry, and characters cannot be separated from the background. When doing camera push-ins, local enlargements, parallax animation, or image-to-video, these flaws will be further amplified.
For film and television creators, the process does not end once an image is generated. It also needs to enter editing, camera movement, visual effects, and animation workflows. This is precisely the meaning of higher quality tiers.
Ace Data Cloud's GPT Image 2.5 has now added two quality tiers, xhigh and max, in addition to low, medium, and high. Both Flare and Sunburst support these two tiers, including the corresponding official channels.
When an image involves complex dark areas, volumetric clouds, smoke, firelight, rainy-night reflections, and multi-layered spatial relationships, creators can let the model invest more generation budget to preserve the structure and details in dark environments as much as possible.
¶ We Tested Five Quality Tiers With a High-Difficulty Dark-Scene Test
To make the differences easier to expose, this time we did not choose a bright background or a simple product, but designed a visual effects keyframe from a science-fiction disaster film.
The scene takes place in a futuristic coastal city at blue hour. A fractured orbital ring hangs above the city, an energy rift tears open the storm clouds, and burning spacecraft wreckage falls toward skyscrapers. In the foreground is a rescue pilot facing away from the camera, with a damaged tilt-rotor aircraft parked nearby.
The entire scene is dominated by dark blue, black, and gray, with only the energy rift, flames, city lights, and parts of the aircraft forming warm highlights.
This kind of scene simultaneously tests whether dark areas contain real details, whether foreground characters can be separated from the background, whether clouds and smoke have spatial layers, whether flames truly illuminate the surrounding environment, and whether the structures of the city, aircraft, and orbital ring are coherent.
The test used gpt-image-2.5-sunburst, which is more oriented toward high fidelity and fine control. All images used exactly the same prompt, with only quality changed.
¶ Full Test Prompt
Cinematic VFX keyframe from an original science-fiction disaster film, ultra-wide 16:9 composition. A colossal orbital ring has fractured above a rain-soaked futuristic coastal megacity at blue hour; a luminous energy rift tears through storm clouds while burning spacecraft debris falls toward skyscrapers. In the foreground, an armored rescue pilot stands on a flooded rooftop beside a damaged tilt-rotor aircraft, seen from behind, looking toward the catastrophe. Include physically plausible volumetric clouds, layered smoke, glowing embers, sparks, rain streaks, wet reflective surfaces, turbulent ocean spray, atmospheric perspective, realistic destruction, miniature-scale city detail, complex blue-orange lighting, accurate reflections and refractions, cinematic anamorphic lens flare, subtle film grain, deep depth staging, photorealistic live-action blockbuster visual effects, coherent geometry, natural human anatomy. No text, no logo, no watermark, no split screen, no collage.
This prompt intentionally includes a large amount of content that can easily lose control in dark scenes: dark clouds and layered smoke, point-like sparks and burning wreckage, dark-armored characters, nighttime city buildings, wet surfaces and complex reflections, as well as multiple light sources with major brightness differences.
If the generation quality is insufficient, the model can easily retain only the intense blue-orange atmosphere while sacrificing the dark-area structure that truly determines the image texture.
¶ Low: The Atmosphere Is Established Quickly, but It Feels More Like an Early Concept Sketch
low completed generation in about 25.11 seconds, with an image output of 157 Tokens.
It can quickly establish a complete scene and can also depict the disaster atmosphere, character placement, and main light sources. It is highly efficient for determining whether a composition works or exploring multiple shot directions in the early stage.
But in complex dark scenes, we cannot only look at whether the image is “dark enough and cool enough.” More importantly, we need to check whether there is still anything in the shadows.
The armor structure of the rescue pilot, the mechanical details of the damaged aircraft, the boundaries between distant buildings, and the brightness variations inside the clouds all determine whether this image can continue to be used in post-production. If this content has already blended together in the original image, simply increasing brightness or sharpening it in post-production cannot truly recover the missing structure.
¶ Medium: Dark Areas Begin to Have More Identifiable Content
medium completed generation in about 27.18 seconds, with an image output of 367 Tokens.
Compared with Low, it does not greatly increase the waiting time, but the model receives more generation budget.
For ordinary video illustration images, Medium is already a relatively practical starting point. It not only needs to complete the overall atmosphere, but also needs to handle the relationships between buildings, vehicles, characters, and the environment more carefully. In dark images, so-called quality improvement does not necessarily appear as “brighter” or “sharper.” More importantly, whether the structures in the shadows can still be seen, whether the smoke has front-to-back layers, whether the distant city can form real atmospheric perspective, and whether the characters naturally separate from the background through rim lighting.
¶ High: Moving from “atmosphere” toward “structure”
high has a generation time of approximately 41.35 seconds, with image output increasing to 1,413 Tokens.
Starting from this level, the model's investment shows a noticeable leap. For cinematic dark scenes, the most noteworthy aspect of this additional investment is the transition between bright and dark areas.
An energy rift cannot be merely a streak of white highlight; it should affect the surrounding clouds. Burning wreckage cannot be merely orange spots; it should produce smoke, embers, and ambient light. Pooled water cannot be merely a simple mirror surface either; it needs to form reflections of different intensities according to the rooftop structure and the direction of the light source.
When these relationships become more complete, the image can gradually move from an atmospheric illustration toward a film shot. High is more suitable for formally entering the video timeline, especially for shots that require subtle push-ins, pans, and depth-of-field processing.
¶ XHigh: Reserving room for enlargement, camera movement, and post-production layering
xhigh completes generation in approximately 53.68 seconds, with image output reaching 2,511 Tokens.
In video workflows, many issues only emerge after the image starts moving. When the camera pushes forward, the distant city needs to withstand enlargement. When creating parallax animation, foreground characters, aircraft, and the background need to have clear boundaries. When handed over to an image-to-video model for continued generation, the mechanical structures and environmental relationships must also be sufficiently clear; otherwise, deformation is very likely to occur during movement.
This is where the value of XHigh lies. It does not merely make the image look more beautiful at first glance, but leaves as much usable information as possible for subsequent production. For dark scenes, this means clearer outlines, more stable local structures, and more layered shadows and smoke.
If the image is to serve as a key shot, an opening visual, a film concept illustration, or the first frame of image-to-video generation, XHigh will better meet actual production needs than ordinary quality levels.
¶ Max: Raising the output ceiling for the most important dark-scene shots
max completes generation in approximately 104.43 seconds, with image output reaching 5,650 Tokens.
Its waiting time is nearly twice that of XHigh, and its image output Tokens are also more than twice those of XHigh. This shows that Max is not merely a name that sounds more advanced, but clearly increases generation investment.
In complex dark scenes, the meaning of the highest quality is not to endlessly add texture, but to preserve as many potentially conflicting goals as possible at the same time: the image must maintain a dark atmosphere, but cannot lose structure; the energy rift must be bright enough, but cannot damage cloud details; the character must be clear, yet cannot look as if they were cut out separately; the smoke must be rich, but cannot obscure all buildings; reflections must be obvious, while also conforming to the ambient lighting.
These requirements are not difficult to achieve individually; the difficulty is making them all hold true at the same time.
Max is more suitable for final output after the composition and prompt have been determined, such as opening key visuals, film posters, core shots, major promotional materials, and images that need to continue into complex post-production workflows.
¶ Quality differences in dark images cannot be judged only by “clarity”
The results of this test are as follows:
| Quality level | Generation time | Image output Tokens | Total Tokens |
|---|---|---|---|
| Low | 25.11 seconds | 157 | 567 |
| Medium | 27.18 seconds | 367 | 777 |
| High | 41.35 seconds | 1,413 | 1,823 |
| XHigh | 53.68 seconds | 2,511 | 2,921 |
| Max | 104.43 seconds | 5,650 | 6,060 |
From Low to Max, what truly needs to be compared is not which image is simply sharper, but how much effective information is retained in the dark environment.
You can focus on observing whether the rescue pilot's armor has recognizable structures, whether there is a clear spatial relationship between the character and the aircraft, whether the aircraft's wings, rotors, and fuselage are coherent, whether the distant city has blurred into one whole mass because of the darkness, whether the clouds have front-to-back depth, and whether the rooftop puddles can reflect multiple light sources in the environment.
What dark scenes fear most is not darkness, but “darkness with no content.”
A high-quality dark image, even if its overall brightness is very low, should still give the characters, buildings, machinery, smoke, and clouds their own distinct layers. Viewers may not explicitly describe these differences, but they will intuitively sense whether the image has a cinematic texture.
¶ Image quality determines how much room remains for video post-production
Video creation rarely uses a completely static original image directly. Creators often use camera push-ins, horizontal movements, local cropping, foreground-background layering, light effect overlays, depth-of-field simulation, or hand images over to image-to-video models for further generation. These operations all amplify flaws in dark areas.
If a character’s outline is already merged with the background, it will be difficult to separate layers when creating parallax; if architectural structures are already blurry, a camera push-in will only reveal larger blur; if smoke has no spatial relationships, it can easily turn into irregular rolling blocks of color when generating dynamic effects.
Therefore, a higher-quality original image is not merely about “making the final image look prettier”; it also directly determines how far post-production can go.
¶ You do not need to use Max for every image, but important dark scenes deserve higher quality
A reasonable production approach is not to set the quality to the highest level from the very first sketch.
During the exploration stage, you can use Low or Medium to quickly confirm the composition, character positions, light source direction, and overall color tone. Once the direction is stable, you can use High to generate formal assets. If the image needs significant enlargement, complex layering, or further video generation, XHigh will be more suitable.
Max should be reserved for the most important shots, such as the first frame of an opening sequence, the main visual of a poster, the core image of a trailer, or key scenes that determine the texture of the entire video.
Ace Data Cloud adds XHigh and Max to GPT Image 2.5. What this truly brings is not two additional parameters, but the ability for creators to no longer remain at the stage of “the atmosphere is close enough” when facing complex dark scenes.
Darkness no longer means giving up detail.
When the model has a higher quality budget, cities, characters, aircraft, smoke, and reflections in the shadows have the opportunity to be preserved together. For film and television creation, this information hidden in dark areas is often precisely the key that takes an image from an ordinary illustration to a cinematic shot.
¶ Experience GPT Image 2.5 Now
Want to use XHigh and Max for your own dark-scene shots, posters, and first video frames? Visit Ace Data Cloud to experience the high-fidelity, finely controlled GPT Image 2.5 Sunburst, or choose the GPT Image 2.5 Flare, which is more geared toward fast image generation.